Fraglingo Generates Molecules with Attachment-Aware Fragment Modeling

Summary

Fraglingo is an autoregressive fragment-based generator designed to reflect common medicinal-chemistry edits, such as extending a scaffold, replacing a substituent, or decorating a molecule at a specified attachment site. Existing methods often select a fragment from a fixed vocabulary and predict its connection separately, which limits generation to closed fragment sets and decouples two related decisions. Fraglingo jointly models fragment identity and attachment in a continuous latent space. It predicts an attachment-aware fragment embedding and retrieves the next fragment through nearest-neighbor search. A wildcard-anchored readout represents the growing molecule from the active attachment site, allowing the embedding to encode both molecular context and the required connection. New fragments can be added to the inference-time vocabulary without retraining when their embeddings can be produced by the trained fragment encoder. The same generation primitive supports molecule generation, scaffold generation, scaffold decoration, and molecular optimization. On controlled property-conditional benchmarks, Fraglingo reports stronger joint property control than comparably trained baselines while retaining competitive validity, uniqueness, and novelty. It also generalizes to fragment libraries up to four times larger than those used during training without retraining.